Abstract
Traditional driving instruction is often limited by instructor availability, unpredictable traffic conditions, and the risks associated with practicing hazardous maneuvers on public roads. This paper presents an AI-driven "Lesson Generator" module for a Virtual Driving Instructor (VDI) that addresses these challenges through procedural content generation. By modeling student proficiency using Bayesian Networks and employing a stochastic utility-based approach to map construction, the system creates bespoke simulator environments tailored to a learner's specific weaknesses. Evaluation using synthetic student profiles confirms that the generator effectively prioritizes high-utility skills, such as tunnel navigation or highway merging, ensuring efficient and targeted practice.
Introduction
Learning to drive is a high-stakes endeavor characterized by significant costs and inconsistent training environments. While simulators offer a safe alternative, many rely on static maps that fail to adapt to individual learner needs. Researchers from the Norwegian University of Science and Technology and Way AS have developed a system that moves beyond these limitations. Their "Lesson Generator" procedurally creates driving lessons that target a student's specific skill gaps in real-time, optimizing the learning curve while maintaining a safe environment.
System Architecture
The VDI framework integrates several sophisticated components to facilitate personalized learning. At its core, the system monitors student behavior through a high-fidelity Unity-based simulator and processes this data to update a dynamic knowledge representation.
The architecture includes several key modules:
- Agent System & Knowledge Graph: Real-time monitoring agents observe student actions (e.g., signaling, mirror checks) and update a Knowledge Graph that maps the current traffic state.
- Student Model: A Bayesian Network that estimates mastery levels for various driving skills based on observations.
- Lesson Generator: The procedural engine that constructs new maps based on the student's current proficiency.
- Traffic Manager: Controls dynamic elements within the VR environment to match the lesson's objectives.
Student Modeling and Skill Mastery
A key innovation of this work is the hierarchical modeling of driving skills. Skills are categorized into "atomic" actions (e.g., activating a blinker) and "composite" maneuvers (e.g., overtaking). By representing these in a Bayesian Network, the system can infer root causes of failure; for instance, a struggle with overtaking might be traced back to poor mirror-checking habits.
The model classifies each skill as Mastered, Learning, or Struggling. These classifications directly inform the "utility" of specific map features during the generation phase, ensuring that the curriculum evolves alongside the student's progress.
Procedural Lesson Generation
The Lesson Generator employs Procedural Content Generation (PCG) to build maps from a library of "tiles" (e.g., T-intersections, roundabouts, highway segments). Rather than a rigid optimization, the system uses a stochastic utility-based approach:
- Utility Calculation: The system identifies skills with the lowest mastery levels and assigns them high utility.
- Tile Selection: Map tiles that facilitate practice of high-utility skills are prioritized.
- Diversity Weighting: To prevent repetitive "blocked" practice, a diversity weight is applied, encouraging "interleaved" practice—a pedagogical technique known to improve long-term retention.
Evaluation and Results
The system was evaluated using four synthetic student profiles ranging from low to high proficiency, including profiles with specific weaknesses (e.g., difficulty with tunnels). The results demonstrated a strong correlation between a student's skill gaps and the content generated. For a profile struggling with highway and tunnel driving, the generator produced maps where these specific tiles appeared with significantly higher frequency. This qualitative success indicates that the system can effectively automate personalized curriculum design.
Conclusion and Future Work
The VDI Lesson Generator represents a significant step toward automated, personalized driver training. While current evaluations rely on synthetic models, future work will involve trials with human learners at the Way AS driving school. Additionally, researchers aim to transition from offline map generation to fully dynamic, "on-the-fly" environment adjustments, further refining the "Tutor" agent's ability to provide nuanced feedback.
Summary updated